2

I have a claim I’ve been conjecturing. Not sure if it’s true or not. Context: I’m doing some calculations with finite difference schemes.

Say I have the following real $n \times n$ tridiagonal matrix $A$:

$$ \begin{bmatrix} \;\;\;2 & -1 & & &\\ -1 & \;\;\;2 & -1 & &\\ & \ddots & \ddots & \ddots\\ & & -1 & \;\;\;2 & -1 \\ & & & -1 & \;\;\;2 \end{bmatrix} $$

Does the following system have a unique solution?

$A\mathbf{U}=\mathbf{F}$ where $\mathbf{U} = \begin{bmatrix} U_{1}\\ U_{2}\\ \vdots\\ \\ U_{n} \end{bmatrix}$ and $\mathbf{F}$ is just a real vector of dimension $n$.

Observations:

Other than the fact it’s tridiagonal, I noticed that it is diagonally dominant. I also think I can compute an $LU$ factorization, but I’m not sure how that would help. Any directions?

Rócherz
  • 4,241
Hinazuki Kayo
  • 63
  • Can you compute $\det A$ ? If $A$ is invertible, then your system has an unique solution. – Dark Jun 22 '16 at 14:11
  • http://math.stackexchange.com/questions/388829/connection-between-irreducibly-diagonally-dominant-matrices-and-positive-definit This question is not quite a duplicate but the desired result follows immediately from the result in this question. – Ian Jun 22 '16 at 14:23
  • This tridiagonal system has a unique solution according to the theorem here –  Jun 22 '16 at 14:27
  • Thanks guys, I see now.

    Could I also do it like this, since $A$ is tridiagonal, an $LU$ factorization exists which can be computed by Thomas algorithm.

    Since $A\mathbf{x}=\mathbf{b}$ iff $(LU)\mathbf{x}=\mathbf{b}$

    We can solve the lower triangular system $L\mathbf{y} = \mathbf{b}$ which yields a unique solution and we can also solve the upper triangular system $U\mathbf{x} = \mathbf{y}$ resulting in $\mathbf{x}$ being unique.

    Also I cannot accept comment answers unfortunately!

     – Hinazuki Kayo Jun 22 '16 at 14:27
  • No, because an LU factorization doesn't guarantee uniqueness. In particular, if $A$ is not invertible then neither is $U$. LU factorization only reduces us to showing that $U$ is invertible. But if, say, you replace the $2$s in the top left and bottom right corner with $1$s, the result is not invertible (constant vectors get mapped to zero; this is a feature shared by all graph Laplacians), yet your logic goes through the same way. – Ian Jun 22 '16 at 14:32

3 Answers3

3

Rodrigo is perfectly right, but we may prove that $A_n$ (the $n\times n$ matrix with such a structure) is invertible by simply computing its determinant through a recursive approach. An expansion along the last row gives: $$\det A_n = 2\cdot\det A_{n-1}-\det A_{n-2} \tag{1}$$ hence: $$ \det A_n = Cn +D \tag{2}$$ and since $\det A_1=2$ and $\det A_2=3$, $\color{red}{\det A_n = n+1\neq 0}$ and $A_n$ is invertible.

$(2)$ follows from the fact that the characteristic polynomial of the sequence $\{\det A_n\}_{n\geq 1}$ is $p(x)=(x-1)^2$ by $(1)$.

Jack D'Aurizio
  • 361,689
2

$\mathrm A$ is not merely tridiagonal, it is also Toeplitz. Hence, the $n$ real eigenvalues of $\mathrm A$ are given by [0]

$$\lambda_k (\mathrm A) = 2 + 2 \cos \left(\frac{k \pi}{n+1}\right)$$

for $k \in \{1,2,\dots,n\}$. Thus,

$$0 < 2 - 2 \cos \left(\frac{\pi}{n+1}\right) \leq \lambda_k (\mathrm A) \leq 2 + 2 \cos \left(\frac{\pi}{n+1}\right) < 4$$

and we conclude that $\mathrm A$ is invertible.

[0] Silvia Noschese, Lionello Pasquini, and Lothar Reichel, Tridiagonal Toeplitz Matrices: Properties and Novel Applications, 2006.

Rodrigo de Azevedo
  • 23,223
0

Note that

$$\mathrm x^T \mathrm A \mathrm x = x_1^2 + \underbrace{\left(\sum_{i=1}^{n-1} (x_{i+1} - x_i)^2\right)}_{=: f(x)} + x_n^2$$

where $f$ is positive semidefinite and vanishes on $\{\gamma 1_n \mid \gamma \in \mathbb R\}$. Fortunately, $x_1^2 + x_n^2$, which is also positive semidefinite, does not vanish on $\{\gamma 1_n \mid \gamma \in \mathbb R\}$, which allows us to conclude that $\mathrm A$ is positive definite and, thus, invertible.

Rodrigo de Azevedo
  • 23,223
  • 1
    It's not quite a Laplacian, because of the first and last rows. The actual Laplacian of a circular graph would have a $-1$ in the top right and bottom left corners. This is important, because graph Laplacians are never invertible. – Ian Jun 23 '16 at 15:24
  • @Ian You're totally right, of course. I will delete that part. – Rodrigo de Azevedo Jun 23 '16 at 15:36